Update app.py
Browse files
app.py
CHANGED
|
@@ -4,10 +4,13 @@ from llama_cpp import Llama
|
|
| 4 |
import requests
|
| 5 |
from tqdm import tqdm
|
| 6 |
|
| 7 |
-
#
|
| 8 |
MODEL_URL = "https://huggingface.co/mradermacher/Saka-14B-GGUF/resolve/main/Saka-14B.Q4_K_M.gguf"
|
| 9 |
MODEL_PATH = "models/Saka-14B.Q4_K_M.gguf"
|
| 10 |
|
|
|
|
|
|
|
|
|
|
| 11 |
def download_model(url=MODEL_URL, path=MODEL_PATH):
|
| 12 |
os.makedirs(os.path.dirname(path), exist_ok=True)
|
| 13 |
if os.path.exists(path):
|
|
@@ -28,36 +31,49 @@ def download_model(url=MODEL_URL, path=MODEL_PATH):
|
|
| 28 |
bar.update(size)
|
| 29 |
print("モデルのダウンロードが完了しました。")
|
| 30 |
|
| 31 |
-
#
|
| 32 |
download_model()
|
| 33 |
|
| 34 |
# モデルロード
|
| 35 |
llm = Llama(model_path=MODEL_PATH)
|
| 36 |
|
| 37 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
response = llm.create_completion(
|
| 39 |
prompt=prompt,
|
| 40 |
temperature=temperature,
|
| 41 |
top_p=top_p,
|
| 42 |
max_tokens=max_tokens,
|
| 43 |
-
stop=["
|
| 44 |
)
|
| 45 |
-
return response
|
| 46 |
|
| 47 |
def chat_interface(user_input, history, temperature, top_p, max_tokens):
|
| 48 |
-
if history is None:
|
| 49 |
history = []
|
| 50 |
-
history.append(
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
|
|
|
|
|
|
| 58 |
|
| 59 |
with gr.Blocks() as demo:
|
| 60 |
-
gr.Markdown("# Saka-14B GGUF
|
| 61 |
chatbot = gr.Chatbot()
|
| 62 |
user_input = gr.Textbox(placeholder="質問をどうぞ", label="あなたの入力")
|
| 63 |
|
|
|
|
| 4 |
import requests
|
| 5 |
from tqdm import tqdm
|
| 6 |
|
| 7 |
+
# モデル情報
|
| 8 |
MODEL_URL = "https://huggingface.co/mradermacher/Saka-14B-GGUF/resolve/main/Saka-14B.Q4_K_M.gguf"
|
| 9 |
MODEL_PATH = "models/Saka-14B.Q4_K_M.gguf"
|
| 10 |
|
| 11 |
+
# システムプロンプト(自由に変更してください)
|
| 12 |
+
SYSTEM_PROMPT = "あなたは丁寧で知的な日本語AIアシスタントです。ユーザーの質問にわかりやすく答えてください。"
|
| 13 |
+
|
| 14 |
def download_model(url=MODEL_URL, path=MODEL_PATH):
|
| 15 |
os.makedirs(os.path.dirname(path), exist_ok=True)
|
| 16 |
if os.path.exists(path):
|
|
|
|
| 31 |
bar.update(size)
|
| 32 |
print("モデルのダウンロードが完了しました。")
|
| 33 |
|
| 34 |
+
# モデルダウンロード
|
| 35 |
download_model()
|
| 36 |
|
| 37 |
# モデルロード
|
| 38 |
llm = Llama(model_path=MODEL_PATH)
|
| 39 |
|
| 40 |
+
def build_prompt(messages):
|
| 41 |
+
prompt = f"<|system|>\n{SYSTEM_PROMPT}\n"
|
| 42 |
+
for msg in messages:
|
| 43 |
+
if msg["role"] == "user":
|
| 44 |
+
prompt += f"<|user|>\n{msg['content']}\n"
|
| 45 |
+
elif msg["role"] == "assistant":
|
| 46 |
+
prompt += f"<|assistant|>\n{msg['content']}\n"
|
| 47 |
+
prompt += "<|assistant|>\n"
|
| 48 |
+
return prompt
|
| 49 |
+
|
| 50 |
+
def generate_response(messages, temperature, top_p, max_tokens):
|
| 51 |
+
prompt = build_prompt(messages)
|
| 52 |
response = llm.create_completion(
|
| 53 |
prompt=prompt,
|
| 54 |
temperature=temperature,
|
| 55 |
top_p=top_p,
|
| 56 |
max_tokens=max_tokens,
|
| 57 |
+
stop=["<|user|>", "<|system|>", "<|assistant|>"]
|
| 58 |
)
|
| 59 |
+
return response["choices"][0]["text"].strip()
|
| 60 |
|
| 61 |
def chat_interface(user_input, history, temperature, top_p, max_tokens):
|
| 62 |
+
if history is None or len(history) == 0:
|
| 63 |
history = []
|
| 64 |
+
history.append({"role": "user", "content": user_input})
|
| 65 |
+
response = generate_response(history, temperature, top_p, max_tokens)
|
| 66 |
+
history.append({"role": "assistant", "content": response})
|
| 67 |
+
|
| 68 |
+
chat_display = []
|
| 69 |
+
for msg in history:
|
| 70 |
+
role = "ユーザー" if msg["role"] == "user" else "AI"
|
| 71 |
+
chat_display.append((role, msg["content"]))
|
| 72 |
+
|
| 73 |
+
return chat_display, history
|
| 74 |
|
| 75 |
with gr.Blocks() as demo:
|
| 76 |
+
gr.Markdown("# Saka-14B GGUF 日本語チャット(システムプロンプト+履歴対応)")
|
| 77 |
chatbot = gr.Chatbot()
|
| 78 |
user_input = gr.Textbox(placeholder="質問をどうぞ", label="あなたの入力")
|
| 79 |
|